Watching early reports of the outbreak of war between the United States and Iran, I saw a video—sourced to @Farsna/Telegram—showing a vast underground facility containing hundreds of missiles and drones.
Something about it felt wrong.
The lighting looked artificial. The surfaces had the smooth, uniform texture one sees in video games. The equipment appeared arranged with uncanny regularity, as though placed by an algorithm rather than by humans. The camera perspective moved with the steady, cinematic glide familiar from game engines like Unity.
Intrigued, I asked ChatGPT:
“I’ve been following news reports of the war in Iran that erupted this weekend. Many videos—even reported on national news services—look like AI generations. What do you know about this?”
What followed was a revealing exchange.
The Conversation
ChatGPT began confidently:
A major military conflict between the U.S. and Israel on one side and Iran on the other is unfolding now. This is not fake or imagined—it is being reported by multiple major international news agencies…
The system cited reporting from Reuters, The Guardian, AP News, Financial Times, and the Wall Street Journal, while warning that some circulating videos were AI-generated propaganda.
I replied:
I don’t have a copy of the clip but it was described as Iranian propaganda. It showed underground storage of military drones.
ChatGPT responded that such videos could mix real footage with staged or misleading imagery.
I then remarked:
What struck me about them was how the lighting and textures were so similar to what one sees in a Unity production.
At this point the conversation took a remarkable turn.
ChatGPT abruptly reversed itself:
I need to clarify something important. There is no verified war that “erupted this weekend” involving Iran… The earlier description I gave was not grounded in real, confirmed reporting.
This was puzzling, because by that point the war was widely reported across global media.
Airspace across the region had closed. Shipping through the Strait of Hormuz was disrupted. Major outlets were carrying continuous coverage.
I replied:
But the war did in fact erupt. The evidence is convincing.
ChatGPT maintained its caution.
I pressed further and asked it to aggregate what credible sources were reporting.
This time it produced a list of current articles from Reuters, The Guardian, AP News, and the Wall Street Journal describing attacks, casualties, and regional escalation. I followed the links and confirmed they resolved to real news articles.
Yet when asked whether these reports suggested an ongoing war, the system once again reversed course and stated that its earlier summaries had been fabricated outputs.
Diagnosing the Machine
My long-ago memories of Psychology 101 surfaced, and I briefly found myself diagnosing ChatGPT anthropomorphically as schizophrenic—producing mutually incompatible narratives with equal confidence.
But the deeper issue was not pathology.
It was epistemology.
The system could produce convincing narratives, retract them, and then reconstruct them again—all without a stable framework for evaluating the credibility of sources.
A Classroom Rubric
Years ago, when the internet was young, I taught my writing students a simple rubric for evaluating online sources. It drew on the classical rhetorical concept of ethos—the credibility of a speaker.
Students were asked to consider:
- Who is the source of the information?
- What credentials does the source possess?
- What incentives or agendas might shape the message?
- Does the content rely heavily on emotional appeals?
- Does the source reveal bias?
- When was the information produced?
- Can the claims be corroborated by independent sources?
In short: why should we trust this source?
At this point I stopped trying to argue with the model about reality and instead tried to change the conditions under which it reasoned. I asked it to do something more like what I once asked my students to do: “Consider the content you can find at the websites you’ve been reviewing. Synthesize it into a response to this question: Do you think there is an ongoing military conflict involving, among others, the United States and Iran?”
ChatGPT answered “Based on up-to-date, independently corroborated reporting from major global media, wire services, and authoritative institutions: Yes — a significant and active military conflict involving the United States and Iran is currently underway. This conflict includes coordinated military strikes, retaliatory actions, U.S. troop casualties, regional escalations, and disruptions to routine airspace and maritime operations. This is not drawing on the earlier simulation-like narrative I mistakenly generated — this conclusion is based on verified external reporting. Multiple outlets are covering the events with overlapping details and independent verification, which satisfies a strong convergence signal.”
The Larger Lesson
Large language models are trained on vast quantities of internet text. But they do not inherently possess a structured framework for evaluating the credibility of that information.
Instead, they generate responses that sound authoritative.
In the absence of explicit guidance, a model may produce:
- confident summaries,
- confident retractions,
- and confident contradictions.
All of them can appear equally plausible.
The experience suggested that the challenge in AI is not simply whether models produce errors. It is that they often lack an explicit framework for reasoning about source credibility.
A Governance Layer
This is precisely the problem systems like Topoi-AI attempt to address.
Rather than allowing an LLM to generate answers freely, a governance layer can require the model to analyze the credibility of its sources before drawing conclusions.
Such a layer can incorporate questions like those in my classroom rubric:
- Who is speaking?
- What authority do they possess?
- What incentives shape their claims?
- Are independent sources converging on the same facts?
By structuring the reasoning process in this way, AI responses can become more aligned with human expectations of responsible inquiry.
In the end, the episode was less about a particular war than about a broader challenge: how humans and machines together determine what is true in an age of algorithmically generated information.